Kelly Criterion Applications in the Cryptocurrency Market

Abstract
The Kelly criterion is a somewhat well-known investment concept in the finance field that
can help investors maximise the expected value of the terminal value of wealth. We
develop a new modified version of a Kelly criterion and test it in a variety of market
conditions and reference time periods. We show that using an exponentially weighted
moving average using a shorter time window, in this case 7 days, investors in the
cryptocurrency market are able to earn returns well above the benchmark; we also show
that in general, investors who are willing to tolerate some level of risk may be able to
outperform in the cryptocurrency market as it likely tends to becomes more normal over
time.

Keywords
Kelly criterion • Cryptocurrency • Exponentially weighted moving average • Exponential utility

1. Introduction
The classical approach to gambling involves a gambler calculating the amount of capital to bet
on each bet in order to maximise total returns in the long run while avoiding the risk of complete
loss. In investing, usage of the strategy known as the Kelly criterion is based on the same
concept – find the fraction of each investment size in each given time period so that the investor
maximizes their total long-term expected returns starting with their given level of capital. The
Kelly criterion in investing terms is typically defined as an edge-over-odds model where an
investor is only recommended to invest should the probability of earning positive return be over
50%, with the proportion of the starting capital to be invested calculated relative to the amount of
excess return, or edge, an investor can expect to receive on a given investment made. As such, a
typical Kelly criterion can be defined as:
𝑏𝑝 − 𝑞
𝑓= (1)
𝑏

Where f is the Kelly fraction (the fraction of available starting capital that should be staked on a
given favorable investment), b is the expected odds received on a given investment, p is the
probability of the investment earning a positive return, and q is the probability of the investment
receiving a negative return. This is a strict representation of a typical edge-over-odds Kelly
criterion.

1
Kelly strategies in general tend to be popular amongst those who use them in both the gambling
and the investing scenarios due to the fact that by definition they can maximise growth of wealth
over time – but there are some clear risks involved. While some, such as Peterson (2018), have
attempted portfolio choice strategies utilising a Kelly criterion, most typical applications do
involve a relatively high level of risk to investors, largely due to a lack of diversification; a Kelly
strategy by its typical definition strictly involves putting some level of available capital into the
favorable asset and the remaining capital into a risk-free asset such as treasury bills.

This paper aims to develop a version of the Kelly criterion that can be used specifically with
more volatile investments, such as cryptocurrencies. As pointed out by MacLean, Thorp, and
Ziemba (2011), fundamentally mathematically a Kelly criterion strategy will maximize the rate
of asset growth (especially when compounding comes into play) and help avoid potential ruin,
even if there are generally more significant risks of losses being undertaken. This paper’s
proposed newer version of a Kelly criterion, utilising a novel way of splitting returns that could
theoretically benefit in a more volatile market (see Section 3), and tested against a relatively
newer market development in cryptocurrency, should help some investors – particularly those
who are more risk-seeking – navigate the market.

By testing this Kelly criterion in a variety of different investment scenarios, time periods,
utilising data over different periods, the optimal scenario or scenarios in which to apply both this
new Kelly criterion and a general version of Kelly criterion relative to the cryptocurrency market
can be found. This paper also aims to compare the Kelly criterion to the use of utility theory to
attempt to maximise expected utility for investors with a spectrum of risk tolerance to see the
potential gains relative to risk trade-off possible for the average investor.

This paper begins with Section 2 providing some background on theory and prior literature
related to the Kelly criterion, cryptocurrency, and utility. The methods and data used for the
study are then described in Section 3, after which the results used from the employment of these
methods and data are described in Section 4. Section 5 is a discussion of the results and the
implications thereof, and the paper ends with conclusion in Section 6.

2. Literature Review
Kelly (1956) developed a strategy for individuals to bet an optimal portion of their capital into
games with favourable odds to generate an exponential growth rate; however, this initial paper was
developed in reference to information rate and information theory rather than thought about in an
investment context. This strategy is known as the Kelly criterion, which will henceforth be referred
to as “Kelly”. The focus of each bet with a Kelly strategy shifts from optimising the expected value
of capital in each period to maximizing the expected value of capital growth. If the winnings from
each bet are reinvested and the gambler retains autonomy over the gambling amount, then the
expected value of the logarithm of capital growth - exponential growth positively or negatively -

2
is accumulated by the gambler. By considering long term growth, the effects of repeated bets is
encompassed in the Kelly formula, thereby encompassing compounding wealth growth.

The Kelly strategy was soon adapted after Kelly’s initial publication, notably by Edward O. Thorp,
and introduced to other games such as blackjack, baccarat, and roulette - most notably blackjack,
where he first popularised the concept of using the Kelly criterion with strategies contained in his
book Beat the Dealer (1962). Unlike the original usage of Kelly in only one-to-one payoff games,
these games did not show such a characteristic and had much higher variation. The typical results
show that a gambler should limit the current maximum bet to a multiple of the minimum bet and
that bet sizes should be adjusted for changes in capital. In later study, Thorp (2008) further
elaborates the use of Kelly in arenas such as sports gambling. The study shows that sports bets are
independent, with the fraction needing to be invested as lower than that of a single bet. The formula
was later modified by Chu, Wu & Swartz (2018) in sports gambling where in determining the
optimal wagering fraction the unknown variable is the probability of placing a winning wager. A
loss function is introduced in order to estimate true Kelly and uses prior distributions as opposed
to frequency distributions. The results obtained show that the original Kelly formula is too volatile,
ergo risky, for most investors, and inconsistent while the modified Kelly led to a criterion based
on minimising expected prior loss and was more consistent.

Kelly has also been adapted and applied to stock markets, though with a requirement that the
discrete probabilities used in the gambling systems were changed to continuous probabilities
(Rotando & Thorp, 1992). Thorp (2008) noted that the situations in which a gambler and investor
operate are analogous and the Kelly criterion could be applied to both. In both cases, participants
aimed to achieve positive returns while incorporating risk into their decisions. This could be
potentially achieved by maximising the expected value of the logarithm of capital as in a Kelly
strategy. Chen & Huang (2008) noted that if investors aim to maximise the expected value of the
logarithm of capital, then they in fact possess logarithmic utility functions. Alternatively, Lo, Orr
& Zhang (2018) proved that if the investor follows expected utility theory, then Kelly is the optimal
strategy for these individuals, even with the inherent riskiness involved. As Kelly almost always
beats other approaches even in different portfolio scenarios, it can help investors allocate wealth
among different assets if adjusted for their level of risk utility. Investors often focus on the rate of
return on the capital they invested and how much their investment grows rather than the Sharpe
ratio of their investment (Estrada, 2009). This is generally the basis on which portfolio
performances are compared.

Utility theory plays a key role in finance, economics, and other fields alike. Expected utility theory
has a long history of use in developing the decision-making process when faced with uncertainty
(Eeckhoudt et al., 2005). The underlying assumptions are that the marginal utility of wealth is
linear but people experience diminishing marginal returns. Scholars for and against the school of
thought have jointly accepted it as the standard theory to use in comparison to other utility theories;
expected utility theory is theoretically sound, thereby making it a good basis for comparison. Some
utility functions, such as constant risk aversion (CARA) utility functions, face criticism about

3
assuming that the risk aversion of an investor does not decrease as wealth increases. However, an
argument can be made that CARA functions make for better comparisons between different assets
classes.

Kuznitz, Kandel & Fos (2008) and Nielsen (2014) further investigated the use of Kelly while
accommodating for the investor’s level of utility while rebalancing the returns gained. The need
for rebalancing is to ensure the original balance is maintained or a new proportion of investment
is made with the returns that were gained. Kuznitz, Kandel & Fos (2008) found that if transaction
costs were ignored, the investor should rebalance when returns are realised. On the contrary, when
transaction costs are involved, rebalancing should only occur on major realised returns or if partial
rebalancing can be used to reduce the impact of transaction costs. The benefit of partial rebalancing
was that it entailed a smoother setting due to its parameters being continuous while intermittent
rebalancing periods were integer based. In Nielsen’s (2014) study, he showed that for risk averse
individuals to optimise they would have to optimise on returns and the power utility function
(constant relative risk aversion). There exists an optimal strategy for investors to invest while
overcoming certain restrictions or constraints - stop-level. Power utility functions are based on two
assumptions (Eeckhoudt et al., 2005) - that the absolute risk aversion is decreasing but risk
aversion relative to level of wealth is constant. These functions negate income effects when an
investor must decide on the proportion of wealth to invest at different levels of wealth while
factoring in risk. This is analogous to the constant proportion of wealth required by the Kelly
Criterion and warrants its inclusion.

Davis & Leo (2011) and MacLean, Thorp & Ziemba (2011) argue the theory building the
foundation of the Kelly criterion is accepted as sound. However, the usage of the Kelly criterion
has not been supported by the majority of scholars and academic institutions due to factors such
as the inherently high level of risk involved (Estrada, 2009). As a result, the mean-variance
criterion is the commonly used and known method of portfolio optimization. The difference is that
Kelly maximizes the growth of the amount of capital invested while the mean-variance criterion
maximizes the expected excess return per unit of risk measured by volatility. In institutions, the
latter is taught with little to no mention made of Kelly (Estrada, 2009). Kelly has not been a popular
choice of theory due to some of its perceived shortcomings. A key assumption of the Kelly
criterion is the emergence of a market with investors that have logarithmic utility. Chen & Huang
(2008) argue that an investor that aims to maximize expected logarithmic utility is not a
conventional investor. However, they concurred that investors are still driven by the same
decisions, although the Kelly criterion for choosing the optimal portfolio has not yet been
established as valid. As there are infinitely many efficient portfolios, Markowitz does not help
identify which is best. Rotando & Thorp (1992), Markusson & Ohlsson (2017) and Chu, Wu &
Swartz (2018) argue that Kelly is highly risky due to the lack of diversification and the need to
heavily re-invest in a consistent asset class. Roll (1973) established that one would not be able to
distinguish between the cross-sectional relations of risk premiums and expected returns of the
Kelly Criterion and the Sharpe-Lintner model. His findings suggested that at the very least the

4
Kelly approach is comparable to the mean-variance approach. He also noted that if the goal of the
investor is to reach a certain terminal value of wealth and time is constant, then an argument for
the superiority of the Kelly Criterion can be made.

Estrada (2009) and Thorp (2011) argue that when certain criteria are met, the Kelly criterion is a
more plausible basis of optimizing portfolios. This contrasts with the more popular mean-variance
basis. The criteria Estrada (2009) proposed are that the investor is relatively less risk averse, the
investor is a long-term investor, and the investor both can and plans on maintaining their long
position holding period. The argument for this choice is that the Kelly Criterion portfolio
maximizes the geometric mean return while its more popular counterpart, the mean-variance
portfolio, maximises the arithmetic mean return. Since geometric better estimates returns when
focusing on the long-run, this would give an inherent advantage to the Kelly strategy. In addition,
when the performance of a portfolio is assessed, investors seldom look at portfolio performance
as a return per unit of risk. Instead the actual return from the portfolio is deemed to be a stronger
sign of performance. This gives more weight to the Kelly Criterion which maximizes growth and
in turn, maximizes returns. Rotando & Thorp (1992), Estrada (2009), Thorp (2011), MacLean et
al. (2011) and Markusson & Ohlsson (2017) all proved that the returns gained through Kelly
significantly outperformed traditional portfolios. In the same time frame, Kelly would be able to
achieve the same result as a traditional portfolios but at a shorter period. However, the optimal
returns gained are much more likely to occur over long-term periods. In the short run it is possible
for Kelly to perform evenly with or even significantly worse than traditional portfolios.

Kelly was first introduced using information theory and proven effective in gambling systems, and
only later was modified and adapted for potential use with investors. With cryptocurrencies
emerging as a relatively attractive new asset class for potential investor gain, further research can
be done in this area. In general, returns are weakly correlated between different cryptocurrencies
and other asset classes (Elendner et al., 2016). This enhances diversification opportunities for
investors. There is downside with cryptocurrency investment, such as high-risk volatility.
However, this can be said to be offset by high expected returns - making cryptocurrencies a
relatively unique asset class by modern standards. An investor who chooses to invest in
cryptocurrency also has the perk of transaction costs being much lower than that of the stock
market. Dyhrberg, Foley & Svec (2018) note that on the most common Bitcoin exchanges, for
instance, average transaction costs can range from 30 basis points to 100 basis points, which is
cheaper than is available with most brokers. Thus, more frequent rebalancing is potentially more
feasible, which could in turn enhance the growth rate of the investment.

The cryptocurrency market shows signs of high volatility, giving it a similar nature to gambling
whilst maintaining the characteristics of a stock market – in that it is traded on exchanges,
historical data is available, and probabilities are continuous (Jiang & Liang, 2017). Similar to that
of Rotando & Thorp (1992), Hagman (2014) and Markusson & Ohlsson (2017) where Kelly had
been used to compare returns against traditional portfolios, this paper shows how the difference in
full and partial Kellys can impact the returns of a portfolio while adjusting for an investor’s given

5
level of utility. Kelly will also be used in conjunction with the exponential utility function to assess
decisions investors would make when faced with investment decisions under uncertainty.
Comparison of the different results will give a clearer image of the type of investor that would
benefit most from use of the Kelly criterion. This paper does not aim to explain different utility
theories but rather, explain the impact they would have when applied to the Kelly criterion when
juxtaposed to the typical “half-Kelly” and “quarter-Kelly” strategies that tend to be used by current
investors as a theoretical way to mitigate inherent risk with the Kelly strategy. In this paper, we
will apply a Kelly and adjusted Kelly to a portfolio holding a cryptocurrency (specifically targeting
Bitcoin) and a safe offset, in this case 1-month US Treasury bills.

3. Data and Methods

3.1 Methods
The traditional Kelly criterion uses an edge-over-odds calculation to determine the allocation an
investor should put into an asset. A Kelly strategy differs from most traditional strategies in that
it tends to be prone to larger exposure to risk while allocating much more of an investor’s wealth
into an asset according to MacLean, Thorp, and Ziemba (2011). Ohlsson and Markusson (2017)
use a Kelly Criterion based on Thorp’s (2006) approximation for a continuous distribution that
we will test for a comparison later in this paper using a risk-free rate and standard deviation
compared to regular returns.

This paper derives a modified Kelly criterion from a paper set forth by Chu, Wu, and Swartz
(2018) that offered a modified Kelly criterion that could be used in sports gambling. Their
proposed criterion was hence modified in order to fit an investment model rather than a gambling
model.

Consider the following. Given that the return of an asset from time period 0 to time period 1 is R,
an investor’s bankroll B1 at time period 1 is calculated assuming a desired Kelly fraction f of their
bankroll B0 at time period 0, as follows:

𝐵1 (𝑓) = (1 − 𝑓 + (1 + 𝑅)𝑓)𝑤 (1 − 𝑓 + (1 − 𝑅)𝑓)1−𝑤 𝐵0 (2)

Where w is a binary variable that is 1 when an asset’s price increases in a given time period, and
0 when it decreases.
More simply, we can define Equation (2) as such:

𝐵1 (𝑓) = (1 + 𝑅𝑓)𝑤 (1 − 𝑅𝑓)1−𝑤 𝐵0 (3)

6
Consider that especially for volatile investments such as cryptocurrencies, a breakdown of
potential investor return into two components might be desired since the empirical distribution
has both skewness and kurtosis that are not captured by an equal sized upward and downward
returns. As such, returns when price increases and returns when price decreases can be split into
two separate variables to measure the level of potential return when prices move in a given
direction. We can therefore amend Equation (3) to reflect this:

𝐵1 (𝑓) = (1+𝑅𝑈 𝑓)𝑤 (1 − 𝑅𝐷 𝑓)1−𝑤 𝐵0

(4)

Now, consider that the ultimate goal for an investor is to maximize the growth in their wealth
over time. To achieve this, we can maximize the expected log growth of wealth over time for an
investor as follows:

𝐵1 (𝑓)
𝐸 [log ] = 𝐸[log((1 + 𝑅𝑈 𝑓)𝑤 (1 − 𝑅𝐷 𝑓)1−𝑤 )]
𝐵0

= 𝐸[𝑤log(1 + 𝑅𝑈 𝑓) + (1 − 𝑤)log(1 − 𝑅𝐷 𝑓)]

= 𝑝log(1 + 𝑅𝑈 𝑓) + (1 − 𝑝)log(1 − 𝑅𝐷 𝑓) (5)

Where p is the probability of an increase in price during the given time period.
Differentiating Equation (5) with respect to f gives us:
𝑝𝑅𝑈 (1 − 𝑝)𝑅𝐷
− =0
1 + 𝑓𝑅𝑈 1 − 𝑓𝑅𝐷

Solving for f:
𝑝(𝑅𝑈 + 𝑅𝐷 ) − 𝑅𝐷
𝑓= (6)
𝑅𝑈 𝑅𝐷

Equation (6) gives another Kelly fraction that can be used to make investment decisions – in
other words, the proportion of an investor’s bankroll in time period n that will be invested until
time period n+1.

In this paper, this modified Kelly fraction is used to make investment decisions over different
investment horizons and using price, return, and empirical probability data calculated from
different periods of time. These results are reflected in the latter sections.

7
For a comparison, we will use a version of a more traditional edge-over-odds Kelly as discussed
in Section 1. Ohlsson and Markusson (2017) propose a relatively simple strategy that is still
mathematically sound and thus will be used for comparison:
𝜇−𝑟
𝑓=
𝜎2 (7)

Where 𝜇 is the mean return, r is the risk-free rate, and 𝜎 2 is the variance of returns.

This paper also examines the usage of an exponential weighted moving average (EWMA) on the
empirical data used in the calculation of the Kelly fraction in order to check if the weighting of
more recent data is an effective strategy for investing in cryptocurrencies using the Kelly method
relative to an equal weighting of all available data. An EWMA is applied across data analysis,
often in finance, as a way to remove noise from data, it could be seen as having the potential to
work to an even greater extent in a newer market such as cryptocurrency. EWMA gives a level
of weighting to all previous data that decreases exponentially, which helps fulfill this removing
noise criterion. An example formula for EWMA, as provided by Raudys and Pabarškaitė (2018)
and used in this paper, can be defined as:
2
𝐸𝑊𝑀𝐴𝑛 (𝑋)𝑖 = 𝛼𝑋𝑖 + (1 − 𝛼)𝐸𝑊𝑀𝐴𝑛 (𝑋)𝑖−1 , where 𝛼 = .
𝑛+1
(8)
Where n is the window of days to be used as a test period for input data for the Kelly formula as
given in Equation (6) and 𝑋𝑖 is the mean of the given data to be tested.

Equation (6) is also tested relative to a strategy that maximized expected utility. Utility is not
commonly used in conjunction with Kelly strategies as there is still no consensus on multi-period
utility optimization. Typically, because Kelly is such a risky strategy, according to MacLean,
Thorp, and Ziemba (2011) many Kelly investors will use a “half-Kelly” or “quarter-Kelly” – i.e.,
calculate the Kelly strategy and halve or quarter it – as a methodology for controlling risk.
However, this method can be thought of as somewhat inadequate given the fact that it is
essentially a linear approximation for utility, and therefore just a risk-neutral assumption.
Therefore, using a proper utility function could potentially be a good approximation for allowing
for an investor’s risk appetite.

An exponential utility equation can be a good approximation for an investor’s required rate of
return given a level of risk appetite in this situation. Mercereau (2004) provides a simple
exponential utility equation, which allows closed-form solutions for potential levels of investor
utility:
−1 −𝐴𝑐
𝑢(𝑐) = 𝑒 (9)
𝐴

8
Where A is a given level of risk tolerance, with values less than 0 representing those who are
risk-seeking and values more than 0 representing those who are risk-averse (risk-neutral
investors, with A of 0, can be ignored); and c representing the item that the utility equation is
seeking to maximise, in this case the growth of wealth over time. Equation (9) will be used to
solve for the Kelly fraction that on a given day maximizes an investor’s expected utility, thereby
yielding an approximate expected return that is more valuable for given levels of investor risk
appetite than just halving or quartering the Kelly result.

3.2 Data
The three cryptocurrencies used to test this modified Kelly criterion in Equation (6) relative to
other methods are Bitcoin, Ethereum, and Litecoin. Bitcoin is an obvious selection for anyone
testing investment in cryptocurrencies as it has by far the largest market cap, trading volume, and
general history of any available cryptocurrencies. Ethereum, as of writing, has the second largest
market cap among all cryptocurrencies and is significantly different enough from Bitcoin to yield
potentially different results due to differences such as its usage of smart contracts rather than a
peer-to-peer system. Ethereum also uses a proof-of-stake system for mining versus Bitcoin’s
proof-of-work, allowing for a higher chance of mining coins based on amount held rather than
just electricity usage. Cocco and Marchesi (2016) define mining as – for instance, relative to
Bitcoin – “a process in which the ‘proof-of-work’ validates a set of transactions and adds them to
the massive and transparent ledger of every past Bitcoin transaction… the generation of Bitcoins
is the reward for the validation process of the transactions.” Proof-of-stake may distort the
market in comparison, however, though Gui, Hortacsu & Tudon (2018) note that it also can
generally protect against attacks due to traceability, leaving some effective differences to
distinguish between the two for investors. Litecoin was an early spinoff of Bitcoin (a so-called
“altcoin”), and represents a coin that is technically effectively the same by technical standards;
however, it spun off early enough to have a large amount of data over time while representing a
significantly high enough market cap over the entire period to be properly differentiated from
Bitcoin.
All cryptocurrency data used is reflected in closing prices – but since trading of cryptocurrencies
technically does not stop at any time, closing prices are defined as the cryptocurrency’s price at
23:59 UTC. Closing prices are used as the smallest time window of investment we assume is an
investment once per day. All cryptocurrency prices are taken from coinmarketcap.com. Data for
daily 1-month US T-Bill yields used in testing of an alternate edge-over-odds Kelly Criterion are
taken from the St. Louis Fed.

4. Results
In this analysis of results, a comparison between different Kelly criteria are used. First, we
compare returns of the modified Kelly derived in Equation (6) to a more traditional edge-over-
odds Kelly and a passive strategy in which the investor buys the cryptocurrency at the beginning
of the tested period and holds until the end. In order to reflect the effects of different rebalancing

9
periods and different sampling period of the empirical distribution, a breakdown of the modified
Kelly into different trading periods – daily, every second day, weekly, and biweekly – as well as
exponentially weighted moving average periods – 7-day, 30-day, 60-day and 90-day - is
depicted. Since many financial models assume that the prices of more recent data are more
relevant, the estimations from the empirical data are then weighted and adjusted to place more
emphasis on recent data. Furthermore, short selling is included into the aforementioned versions
of the modified Kelly tests to allow for a more generalized test of the strategy. The use of short
selling is a possible option for investors in the crypto space, at least for the cryptocurrencies
tested in this paper. Even though all tested cryptocurrencies have increased significantly over
time, there are incredibly bearish periods in which investors may be able to take advantage of a
downswing quickly before moving out of the position. In addition, exponential utility is tested to
see if an alternative to the typical linear strategies that lower risk from a Kelly investment is
feasible in the context of the Modified Kelly set out in this paper.

It should be noted that the following analysis included the following assumptions:
 The investment window used is 1 April 2017 to 31 August 2018, though other potential
windows are included in one part of the section. This is a relatively short period to test for
a Kelly strategy due to the general aim to maximise long-term capital growth but since
cryptocurrency markets have been relatively underdeveloped until recently – for instance,
Ethereum was only founded in mid-2015 - this period is the only where there is data for
all the cryptocurrencies in consideration while leaving enough data for the estimation of
the empirical distribution. It should also be noted that this investment period incorporates
a “pre-boom” period (approximately from April 2017 until November 2017), the “boom”
period (approximately December 2017), and the “post-boom” period (approximately
January 2018-present), as will be referred to later in this paper.
 The values for probability of a positive return and the values for the predicted up-return
and down-returns were estimated using the empirical distribution. The number of
historical days used depended on the given strategy.
 We assume no transaction costs. This includes all fees which would be included in
transferring funds from platform to platform and the fees charged for short selling, as
well as taxes.
 This study does not include leveraging. Therefore, when a Kelly allocation indicates that
an investor should use an allocation in the tested asset of greater than one, which
indicates that the investor should take on leverage in order to maximise long-terms
returns, it is capped at one.
 When the Kelly fraction is less than one, it is assumed that the investor holds their
remaining wealth as cash rather than investing in a risk-free asset due to the negligible
impact this has on comparative returns.

10
  When short selling was not used, if the Kelly fraction was less than zero it was capped at
zero; if short selling was used, if the Kelly fraction was less than -1 it was capped at -1 to
ensure the lack of leverage trading.
 It is assumed that all investment decisions and subsequent investments are made at 23:59
UTC once the “closing price” data for the day becomes available. All transactions are
done through an exchange from US Dollars to cryptocurrency or vice versa.
 It is generally assumed for the purposes of this paper that any given potential Kelly
investor, to some extent, is risk-seeking, though this does not necessarily have to be the
case. Many of the recommended strategies that would maximise investor wealth are
inherently very risky – sometimes requiring a 100% investment in only one asset – and
investors who are risk-averse are unlikely to be willing to commit to such a strategy.

4.1 Analysis of the Modified Kelly versus Traditional Kelly and Passive

The strategies were compared over the time period 1 April 2017 to 31 August 2018, henceforth
referred to as the “Main Time Period.” The Modified Kelly strategy calculates the appropriate
allocation of wealth based off of three variables – the probability of the price increasing, returns
when prices increased, and returns when prices decreased. The investment is assumed to be
rebalanced daily to ensure returns are optimised. The traditional Kelly, as shown in Equation (7),
differs in investment allocation as it uses an edge-over-odds approach. Both strategies use
historical empirical data to infer conclusions as to how current decisions should be made. Both
strategies are additionally compared to the benchmark which is a passive buy-and-hold strategy.
The following is an illustration of the methodology used for each strategy.

On 1 April 2017, the investor starts with an initial capital of $1000. Using all of the historical
data prior to 1 April 2017, we estimate the required parameters for each of the strategies and
calculate the associated fraction the investor would allocate to the strategy. We then move the
clock forward and calculate the wealth an investor would have based on the change in price of
the underlying cryptocurrency on 2 April 2017. The process is then repeated to calculate the new
fraction an investor should allocate to each strategy using all of the data prior to 2 April 2017.
The whole process is repeated until the clock reaches the final date of 31 August 2018.

Table 1 provides a summary of the overall returns generated over the time period 1 April 2017 to
31 August 2018.

Table 1 Mean returns from different methods

1/04/2017 - 31/08/2018
Modified Traditional Passive
Kelly Kelly Investment

11
 Bitcoin 551,33% 0,44% 551,33%
Ethereum 458,19% 743,00% 458,19%
Litecoin 800,15% 226,02% 800,15%

The results do not follow a similar trend to prior studies where Kelly strategies outperforms the
passive investment strategy. The overall returns generated using the modified Kelly are the same
as the passive investment strategy because the Kelly allocation in the modified strategy suggests
to the investors to invest their entire wealth in each cryptocurrency for every day over the period,
effectively acting the same as a passive strategy would. However, it should be noted that the
recommended Kelly allocation for both Bitcoin and Litecoin began higher but continued a
downward trend towards the end of the period, suggesting that as more data is accumulated a
more realistic Kelly allocation not investing full wealth into the asset may occur – but more
research is likely needed. It can be reasonably assumed that the allocation would normalise over
time as the market for cryptocurrencies begins trending more towards a less volatile distribution
as has occurred in recent months. Additionally, both Modified Kelly and the passive strategy
perform better than Traditional Kelly with the exception of Ethereum. This is potentially an
anomaly due to Ethereum having less available prior data than the other two coins.

4.2 Analysis of Modified Kelly adjusted for different trading periods

In this section, the effects of rebalancing periods on the modified Kelly investment on the Kelly
portfolio are tested by rebalancing every second day, weekly, and biweekly periods. Investing
through these periods would reduce the amount of times rebalancing would occur which does
have the benefit of ultimately reducing transaction costs. The Main Time Period was also used
for this portion of study, but precise starting and ending dates may vary based on necessity
within the time period; for instance, the biweekly trading period ends on 22 August 2018 as the
next possible trading date falls outside of the Main Time Period. The methodology illustrated
above is the same methodology employed in this section, but in each step the clock is moved
forward by the appropriate rebalancing period.

The mean and the standard deviation in Table 2 represent the daily mean and volatility of the
experienced returns for each of the strategies
Table 2 Descriptive Statistics for different trading periods
1/04/2017 - 31/08/2018
Daily Alternate Day Weekly Biweekly
Bitcoin
Total Return 551,33% 545,83% 176,86% 176,86%
Mean 0,48% 1,00% 1,39% 4,73%
Standard Deviation 4,95% 7,54% 13,94% 20,18%
Ethereum
Total Return 458,19% 460,37% 451,58% 483,45%

12
 Mean 0,53% 1,12% 4,06% 8,89%
Standard Deviation 6,42% 9,70% 19,28% 30,93%
Litecoin
Total Return 800,15% 776,74% 483,77% 732,56%
Mean 0,71% 1,40% 4,61% 10,85%
Standard Deviation 7,76% 10,93% 22,97% 33,90%

As per Table 2, all investment strategies involving daily trading outperforms the alternate
strategies except for Ethereum where investing in alternate days and biweekly outperform. This
outperformance in Ethereum is the result of the periods optimising price changes and volatility –
more days with large price gains. This is then classified as an anomaly as the trading period is
decided by the investor and will not always be able to optimise on upward price trends. While
biweekly returns may be higher occasionally, it is also consistently the strategy that exposes the
investor to the most volatility. Bitcoin and Litecoin give optimal returns regardless of risk
adjustments if an investor follows a daily trading period.

4.3 Exponentially Weighted Moving Average

The estimation of the parameters for each of the Kelly fractions (equations 6) are calculated from
the empirical distribution. In order to improve the quality of the estimated parameters, we
examine the use of an exponentially weighted moving average on each of the input data streams.
Therefore, to obtain a value of f for each time period, the exponentially moving average formula
from Equation (8) was applied to each of the inputs of Equation (6). It is worth noting that from
Equation (8) α is the level of weighting of the most recent data. As α = 2 / n+1, it decreases in
magnitude as the time period, n, increases.
The effectiveness of 4 different exponential weighting periods - 7 days, 30 days, 60 days and 90
days – were investigated. A shorter weighting period is in general a more reactionary strategy.
Recent data is factored in more quickly and weighted more heavily compared to the average of
the previous data, and thus investment decisions are more in line with the most recent market
sentiment. It is easy to see the potential of this strategy, therefore, in more volatile markets such
as cryptocurrency markets, especially when bullish and bearish trends are common, as “riding
the wave” has the potential to lead to larger success.

The following charts illustrate the returns achieved by the strategy using the different weighting
periods for each of the different cryptocurrencies.

13
Fig. 1 Bitcoin: EWMA Cumulative Returns

Figure 1 is a plot of the cumulative returns of Bitcoin. The results showed total cumulative
returns of 10232,37%, 2119,44%, 1250,69% and 876,78% for the 7-day, 30-day, 60-day and 90-
day periods respectively. The 7-day period had the largest cumulative return and it was traded
throughout the period due to larger fluctuations in data from which the Kelly fraction was
calculated. Due to the larger α of 0.25, opportunities to benefit from the short term bullish and
bearish trends that have tended to happen with Bitcoin – though mainly bullish – tend to occur
more frequently. The 30-day and 60-day periods as shown have a lower α and require a longer
bullish market run to recommend any significant allocation in Bitcoin. When this requirement
was not met, the Kelly fraction decreased to zero, the investment strategy became hold and
cumulative returns were capped. The 30-day period had a higher cumulative return than the 60-
day return which in turn had a higher cumulative return than the 90-day period with α90 – the
window with the lowest cumulative return due to the longest overall recommend periods of
holding.
Fig. 2 Ethereum: EWMA Cumulative Returns

14
Figure 2 is a plot of the percentage cumulative returns of Ethereum using EWMA. The results
showed a cumulative return of 27758,80%, 6197,83%, 1458,37% and 1666,98% for the 7-day,
30-day, 60-day and 90-day periods respectively. The 7-day period also continued to trade
throughout the period for reason similar to those of the 7-day period of Bitcoin in that the
volatility in data inputs contributed to wild swings in recommended Kelly allocations. The
results generally followed the general trend followed by Bitcoin EWMA cumulative returns with
the 30-day period generating the second highest returns while also switching between trade and
hold strategies depending on the length of the historic bullish run in the market. In the case of
Ethereum the 90-day cumulative returns were higher than the 60-day returns; however, this can
be chalked up to luck due to the relatively small difference between the two returns.

As was the case with Bitcoin, all the periods using exponentially weighted moving average
maximise the possible daily return and cap the maximum losses. Maximum returns across daily
returns due to price change was 33,66% and while 60-day and 90-day periods held maximum
losses of 22,81%, the 7-day and 30-day periods capped them to 17,81% and 18,44%
respectively.
The mean daily returns for the 7-day, 30-day, 60-day and 90-day period of Ethereum were
1,21%, 0,92%, 0,69% and 0,71% respectively, as compared to 0,53% from strictly holding. The
weighted average strategy across all periods outperformed the same strategy relative to Bitcoin
on average, though this can be partially attributed to Ethereum seeing a larger percentage
increase in price over the Main Time Period. Once again, the 7-day window for EWMA held the
largest returns and lowest standard deviation.
Fig. 3 Litecoin: EWMA Cumulative Returns

15
Figure 3 is a plot of the cumulative returns of Litecoin. Like Ethereum, Litecoin was in general
more volatile than Bitcoin over the Main Time Period. The results were similar to the Ethereum
results. The cumulative return for the 7-day, 30-day, 60-day and 90-day period were 31790,77%,
7007,76%, 2233,48% and 3010,30% respectively. The relative performance of returns was also
similar to those of Ethereum and the general extra risk from the increase in volatility was
compensated through increase in returns. It should be noted that since the 90-day window
outperformed the 60-day window in returns for both Ethereum and Litecoin, which are in general
more volatile, that the 90-day window can be said to be preferable for more volatile
cryptocurrencies.

As was the case with Ethereum and Bitcoin, all the periods using exponentially weighted moving
average maximise the possible daily return and cap the maximum losses. Maximum return across
daily returns was 47,60% and while 60-day and 90-day periods had maximum loss of 32,64%,
the same as the maximum loss due to price change, the 7-day and 30-day periods capped them to
15,14% and 29,60% respectively. This finding is once again significant in favor of the 7-day
window; the potential ruin is capped significantly in an even more volatile market than Bitcoin.

The average daily return for the 7-day, 30-day, 60-day and 90-day mean for the daily returns of
Litecoin were 1,30%, 1,04%, 0,83% and 0,91% respectively, as compared to a 0,73% return
from strictly holding. The weighted average strategy across all periods outperformed the holding
Bitcoin on average. The standard deviation for the same periods for Kelly daily returns were
6,33%, 6,85%, 6,84% and 7,22% which was lower than that of the daily price returns of 7,76%.

16
 Table 3 Rolling Moving Average Returns
1/04/2017 - 31/08/2018
Returns – No short selling
7 Day (%) 30 Day (%) 60 Day (%) 90 Day (%)
Bitcoin 27329,48 2396,44 1359,70 900,33
Ethereum 13080,31 11517,75 10322,27 10322,27
Litecoin 100725,97 7689,94 2325,16 3021,87

Returns – With short selling

7 Day (%) 30 Day (%) 60 Day (%) 90 Day (%)
Bitcoin 86805,04 5657,75 1625,56 978,20
Ethereum 842315,25 37757,15 3913,27 3366,10
Litecoin 267401,52 28637,33 4361,37 6625,78

Table 3 then shows the descriptive statistic of the daily returns of price and for each of the
exponentially weighted moving averages. The daily return for price is the percentage increase or
decrease in price from the price of the prior day. The daily return is the percentage increase in
wealth using the specific weighting period model. The output was obtained through data analysis
tools in Excel.

Table 4 EWMA Descriptive Statistics – No Short Selling

Bitcoin - Regular

Rolling window in days 1/04/2017 - 31/08/2018

Price (%) 7 Day (%) 30 Day (%) 60 Day (%) 90 Day (%)

Overall Return 551,33% 10232,37% 2119,44% 1250,69% 876,78%

Mean 0,48% 0,96% 0,68% 0,58% 0,52%

Standard Deviation 4,95% 3,53% 3,99% 3,95% 4,10%

Max 25,25% 25,25% 25,25% 25,25% 25,25%

Min -18,74% -9,78% -18,26% -18,74% -18,74%

Skewness 0,38 1,80 1,16 1,14 0,89

Kurtosis 3,13 9,65 7,79 8,50 7,80

Ethereum - Regular

Rolling window in days

Price (%) 7 Day (%) 30 Day (%) 60 Day (%) 90 Day (%)

17
 Overall Return 458,19% 27758,80% 6197,83% 1458,37% 1666,98%

Mean 0,53% 1,21% 0,92% 0,69% 0,71%

Standard Deviation 6,42% 4,86% 4,91% 5,68% 5,72%

Max 33,66% 33,66% 33,66% 33,66% 33,66%

Min -22,81% -17,81% -18,44% -22,81% -22,81%

Skewness 0,75 1,97 1,70 1,19 1,11

Kurtosis 3,11 8,89 8,82 5,96 5,91

Litecoin - Regular

Rolling window in days

Price (%) 7 Day (%) 30 Day (%) 60 Day (%) 90 Day (%)

Overall Return 800,15% 31790,77% 7007,76% 2233,48% 3010,30%

Mean 0,71% 1,30% 1,04% 0,83% 0,91%

Standard Deviation 7,76% 6,33% 6,85% 6,84% 7,22%

Max 47,60% 47,60% 47,60% 47,60% 47,60%

Min -32,64% -15,14% -29,60% -32,64% -32,64%

Skewness 1,47 2,98 2,14 1,94 1,82

Kurtosis 6,92 15,01 11,75 11,61 9,80

All the periods in Table 4 tend towards the largest possible daily return. This is because when
market sentiment is bullish – the Kelly fraction trends towards 1; and the increase in the portfolio
is close to or exactly one-to-one with the increase in the price of the underlying. Thus, the
maximum daily return experienced in the period by the price and across all weighting periods
was 25,25%. Another added benefit is that when market sentiment is bearish, the Kelly fraction
tends to decrease towards and past -1. This lowers the percentage of the portfolio invested and
minimizes potential ruin from a decrease in the price of the underlying. While 60-day and 90-
days periods had a maximum loss of 18,74%, the 7-day and 30-day periods capped losses at
8,78% and 18,26% respectively. This is obviously much more notable for the 7-day window, and
potentially significant in terms of minimising relative potential investor ruin.

To gain a better understanding of the performance of the exponentially weighted moving average
portfolios, the return relative to risk was also considered. From table 3, the mean daily returns for
the 7-day, 30-day, 60-day, and 90-day windows for Bitcoin were 1,15%, 0,70%, 0,60%, and
0,53% respectively, as opposed to 0,48% from strictly holding. The weighted average strategy

18
across all weighting periods outperformed the holding Bitcoin on average. The standard
deviation for the same periods was 3,60%, 3,97%, 3,94% and 4,09% for the 7-day, 30-day, 60-
day and 90-day periods. The standard deviation across all the weighted periods is less than that
of the daily price returns, 4,95%, across all weighting periods, a positive sign for the mitigating
risk factor of the strategy. The 7-day window also outperforms all others in this aspect with the
highest returns and lowest volatility.

These findings indicate that a 7-day window used in conjunction with an exponentially weighted
moving average may be a significant way to generate outperforming returns for investors.
Among all versions of the strategy tested and as compared to the benchmark, the 7-day window
EWMA consistently generated the highest mean and cumulative returns, both daily and total,
while also having the lowest volatility and capping potential investor ruin much more
significantly than any of the other strategies.

When short selling is taken into account (see Table 5) potential investor returns increase
significantly, especially in regard to the 7-day window. For the 7-day window, potential investor
returns increase 9-fold for Bitcoin, and this is on the low end; potential investor returns for
Litecoin increase 11-fold and for Ethereum the increase is 25-fold. There are somewhat
significant increases for the 30-day window, the 60-day window sees smaller increases, and the
90-day windows sees increases for Ethereum and Litecoin but a decrease in Bitcoin; this might
be due to the lower overall allocation the Modified Kelly would suggest in Bitcoin during
segments of the rolling window.

This continues to support a 7-day rolling window as a potential investment strategy. Although
shorting is somewhat difficult for non-Bitcoin cryptocurrencies at present moment, taking into
account recent information seems to boost significantly potential returns; booms and busts are
effectively taken advantage of, even outside of the December-January boom-bust period that
increases the seeming potentially incredible returns shown. The downside protection that the 7-
day window provides also continues to be evident.

Table 5 Descriptive Statistics of Returns - Short Selling

Bitcoin - with Short selling
Rolling window in days 1/04/2017 - 31/08/2018
Price (%) 7 Day (%) 30 Day (%) 60 Day (%) 90 Day (%)
Overall Return 551,33% 86805,04% 5657,75% 1625,56% 978,20%
Mean 0,48% 1,42% 0,90% 0,66% 0,57%
Standard Deviation 4,95% 4,62% 4,70% 4,71% 4,66%
Max 25,25% 25,25% 25,25% 25,25% 25,25%
Min -18,74% -15,40% -18,27% -18,74% -18,74%
Skewness 0,38 0,61 0,72 0,60 0,53
Kurtosis 3,13 3,67 3,69 3,73 4,02
Ethereum - with Short selling

19
 Rolling window in days
Price (%) 7 Day (%) 30 Day (%) 60 Day (%) 90 Day (%)
Overall Return 458,19% 842312,25% 37757,15% 3913,27% 3366,18%
Mean 0,53% 1,94% 1,32% 0,90% 0,88%
Standard Deviation 6,42% 5,99% 5,84% 6,14% 6,34%
Max 33,66% 33,66% 33,66% 33,66% 33,66%
Min -22,81% -17,81% -18,44% -22,81% -22,81%
Skewness 0,75 0,82 0,84 0,86 0,73
Kurtosis 3,11 3,23 3,60 3,70 3,17
Litecoin - with Short selling
Rolling window in days
Price (%) 7 Day (%) 30 Day (%) 60 Day (%) 90 Day (%)
Overall Return 800,15% 267401,52% 28637,33% 4361,37% 6625,78%
Mean 0,71% 1,79% 1,36% 0,99% 1,09%
Standard Deviation 7,76% 7,38% 7,47% 7,33% 7,68%
Max 47,60% 47,60% 47,60% 47,60% 47,60%
Min -32,64% -21,97% -29,60% -32,64% -32,64%
Skewness 1,47 1,74 1,49 1,50 1,43
Kurtosis 6,92 7,58 7,57 8,11 6,99

4.4 Exponential Utility

In this instance the fraction of the investor’s wealth to invest in the different cryptocurrencies is
determined by maximizing their expected utility. Utility in this case is an exponential utility as
defined in Equation (9) with different risk aversion parameters. Since the actual investor’s risk
aversion parameter is unknown, we estimate the efficient frontier by using 5 equally spaced risk
aversion parameters i.e. 1, 0.5, 0.25, -0.25, -0.5 and -1.0.
The following is an illustration of the methodology used for this strategy:
On 1 April 2017, the investor starts with an initial capital of $1000. Given a risk aversion
parameter and an investment horizon, a given investor’s expected utility is maximised. The
expected utility is calculated using Monte Carlo simulation that randomly selects a date prior to 1
April 2017 minus the investment horizon. For example, if the investment horizon is 20 days then
the dates selected would occur prior to 11 March 2017. For the chosen date, the utility is
calculated using Equation (9) where the risk aversion parameter is given and the terminal wealth
is calculated using the return from the selected date to the selected date plus the investment
horizon. In order to ensure convergence of the mean utility, 20000 dates are randomly simulated
with replacement. The maximization is performed by stepping over all fractional investments
from 0% to 100% in steps of 1% and simulating the expected utility for each fraction. The
fraction that maximises the expected utility is then used as the investment fraction for 1 April
2017. The clock is then moved forward, and the wealth an investor would have based on the

20
change in price of the underlying cryptocurrency on 2 April 2017 is calculated. The process is
then repeated to calculate the new fraction an investor should allocate to the strategy using all of
the data prior to 2 April 2017 minus the investment horizon. The whole process is repeated until
the clock reaches the final date of 31 August 2018.

The above process is then repeated 5 separate times at the same risk aversion parameter and
investment horizon to check for stability of this method and ensure that the outcomes are
consistent. If the percent change in the outcomes is too large, then it can be concluded that
maximising expected utility is too sensitive to the input parameters and therefore not a viable
investment methodology for cryptocurrencies.
Finally, all of the above processes are performed for each of the 6 risk aversion parameters – the
more negative a risk tolerance, the more risk-seeking an investor; and the more positive, the
more risk-averse. Additionally, potential short selling is ignored.

It should be noted that Bitcoin and Litecoin had 1952 total price data points available and
Ethereum had 1121 total price data points available.

Table 6 Average Exponential Utility Results

Risk Tolerance Net Return Daily Mean Daily Standard Deviation

Bitcoin
-1 675,37% 0,51% 4,80%
-0.5 549,68% 0,48% 4,94%
-0.25 549,54% 0.48% 4,94%
0.25 142,10% 0,10% 1,48%
0.5 65,36% 0,10% 0,97%
1 39,73% 0,07% 0,68%
Ethereum
-1 730,19% 0,58% 5,94%
-0.5 458,37% 0,53% 6,42%
-0.25 457,08% 0,53% 6,42%
0.25 128,92% 0,17% 1,68%
0.5 67,12% 0,11% 1,07%
1 31,45% 0,05% 0,63%
Litecoin
-1 1061,41% 0,56% 4,29%
-0.5 1538,72% 0,69% 5,52%
-0.25 1711,20% 0,76% 6,44%
0.25 54,38% 0,09% 1,10%
0.5 31,48% 0,05% 0,65%
1 6,18% 0,01% 0,34%

21
Using an investment horizon of 30 days, the results for each of the risk aversion parameters for
each of the cryptocurrencies is listed in Table 6. Comparing these strategies to the passive
strategy, Litecoin for investors who are risk-seeking tended to show outperforming returns for
the Main Time Period compared to the 800,15% achieved over the period for the passive
strategy. Otherwise, returns ranged from barely outperforming a passive strategy for risk-seeking
investors to significant underperformance for those who are more risk-averse. Overall, this
strategy did not show much outperformance through an exponential utility-based strategy outside
of risk-seeking investors in Litecoin.

However, since the distribution of the sample mean utility should have an error proportional to
1/√20000, we would expect the resulting terminal wealth to be approximately the same for each
trial undertaken. The results approximately converge at a 2% tolerance level for those who are
risk-seeking, while convergence for risk-averse investors is somewhat less stable at around a 5-
10% tolerance level before convergence. This proves that to some extent, an exponential utility
may be a viable risk dampener that helps investors pick an appropriate Kelly allocation as
opposed to halving or quartering a resulting Kelly fraction, assuming that the investors is willing
to take on the risk inherent in a high Kelly allocation in a single asset. However, it is prudent to
conclude that maximizing expected utility for cryptocurrencies is likely not an incredibly tenable
strategy when, outside of Litecoin investors, returns are mostly comparable to investors who
undertake a passive strategy when using a utility strategy often frequent rebalancing of the
portfolio – often on a daily basis - thereby incurring significant transaction costs.

One major contrast inherent to this type of strategy is that a Kelly strategy is more useful as a
multi-period model whereas utility is more of a single-period model. When seeking to maximise
expected utility, the investor cares only about immediate growth in the short-term, whereas when
executing a Kelly strategy the investor seeks to maximise long-term expected growth of their
wealth. Therefore, it is possible that the two strategies cannot fully coalesce, partially explaining
the overall lackluster return outcomes shown in Table 6.

4.5 Post-Boom Analysis

Much of the discussion and analysis in this paper revolves around a defined Main Time Period
that involves a large boom in the cryptocurrency market at around 6 months in that may be said
to amplify compounding returns – as Kelly strategies are wont to do. As such, evaluating
potential returns using the above outlined strategies during a period in which most
cryptocurrencies, including the three tested in this paper, experienced an overall decline and
stabilisation, can give a clearer picture as to whether the modified Kelly criterion as defined in
Equation (6) can add value for investors. The results of this testing can be seen in Table 7,
indicating that the Modified Kelly strategy is still viable for cryptocurrencies when no significant
boom or increase period occurs.

22
 Table 7 Post-Boom Analysis
1/02/2018-
31/08/2018
Bitcoin Ethereum Litecoin
Passive -23,26% -72,70% -56,48%
Modified Kelly
Daily -23,26% -72,70% -56,48%
Every Other Day -23,91% -72,60% -57,61%
Weekly -26,29% -70,99% -63,86%
Biweekly -29,99% -69,31% -64,26%
EWMA
7 Day 233,21% 154,28% 149,17%
30 Day 70,26% 76,69% 76,60%
60 Day 5,92% -23,79% 0,13%
90 Day -1,87% -13,59% 50,38%
EWMA – Short Selling
7 Day 1114,05% 1874,80% 774,67%
30 Day 206,65% 624,06% 351,02%
60 Day 12,13% 96,25% 91,44%
90 Day 8,32% 69,50% 225,18%

The results support the results obtained over the initial investing period. The passive investment
strategy produces returns that are comparable to the Modified Kelly daily, every other day,
weekly, and bi-weekly, similarly to the initial results. This is due to the constant recommended
Kelly allocation above 1 and lack of leverage used.
The exponentially weighted moving average continues to outperform the passive strategy with
shorter weighting periods performing better than longer periods. Again, the 7-day exponentially
weighted period performed better than all other Kelly strategies producing returns of 233,21%,
154,28%, 149,17% with no short selling for Bitcoin, Ethereum and Litecoin respectively.
Allowing for short selling, returns improve to 1114,05%, 1874,80% and 774,67%. In the same
time frame, investing in the underlying alone – the passive strategy, would have produced losses
of -23,26%, -72,70% and -56,48%. Table 7 supports the claim from evidence previously stated in
results, in the cryptocurrency market, there is value to be added from following the modified
Kelly criterion and a 7-day EWMA. Even when market sentiment was bearish, this was reflected
in the Kelly fraction of less than zero which meant short selling and overall returns that
surpassed other strategies.

23
5. Discussion
The cryptocurrency market is a new and emerging market where, now that the market boom of
December 2017 has passed and the market seems to have begun to normalize, investors may be
able to take advantage of a stable yet higher volatility market. The results clearly show that
capital invested in the cryptocurrency market, irrespective of strategy, will have produced high
return over time assuming an investor was willing to take on some level of additional risk. There
are potential further decisions to be made – with the recommended Kelly allocation often larger
than 1 or lower than -1, the investor would theoretically maximize the growth of their log wealth
through taking on leverage – possibly including short selling - and with the restrictions
undertaken in this study, there is potential for even larger returns for the investor willing to take
on further risk than has already been recommended. The key finding in this paper were the
monumental returns that were generated using a 7-day rolling period with exponentially
weighted moving averages, including short-selling. The modified Kelly adapted to all of the
rolling periods, but most specifically the 7-day window, with this window having significantly
higher returns and lower volatility and potential ruin than what would be available either from
holding in the cryptocurrency market or holding in a regular market such as the S&P 500 (for
comparison’s sake, the growth of a typical S&P ETF over the Main Time Period was
approximately 27%). Short selling would be recommended if an investor can afford the
significant extra burden of the transaction costs that would be involved in a daily rebalancing of
a shorted cryptocurrency portfolio. Leverage is another method investors could use to unlock
greater returns – especially with the often bullish trends that cryptocurrency markets have
experienced, a quick leverage in an uptick period could result in greatly magnified returns.
Finally, the exponential utility function was used to adjust potential investor returns with respect
to their level of risk tolerance. Generally, the more willing the investor is willing to take on risk,
the more returns could be expected from the market, although in most cases with usage of the
exponential utility function the returns are likely not worth the constant rebalancing and
subsequent transaction costs involved in the strategy.

This paper discusses the use of Kelly in a cryptocurrency market and the utility that an investor
will derive. In general, prior applications of Kelly criteria have focused on gambling, sports
betting, and stock markets. As mentioned earlier, cryptocurrencies are a relatively new market
that tend to promise investors high returns, though over the past few months the market has
become more normalized while still containing significant volatility. This paper proposes that
using a modified Kelly adapted to the optimal rolling period of 7-days with exponentially
weighted moving averages is the best method to outperform a passive strategy in the
cryptocurrency market. The 7-day window is ultimately most optimal as it offers investors
enough flexibility to take advantage of immediate price fluctuations while the tested other
periods, due to the time duration, are less flexible. The 7-day window also contains enough data
to feel comfortable accepting that a trend is occurring as compared to even smaller potential
windows, which may not contain enough data to establish a trend – or may bypass enough data

24
points to miss a trend. Essentially, the 7-day rolling period allows the investor to make the most
optimal prediction in the market, while even offering a much lower volatility and potential ruin.
However, even while the 60-day and 90-day may offer returns which are comparably lower to
the other rolling periods, they are still higher than those achieved through a passive investment
strategy – but since the 7-day window is more optimal in every conceivable way than the other
tested windows, it is the only one that can be recommended as an option. It can be theorised that
in general this is due to the nature of the cryptocurrency market up until present. Cryptocurrency
investors have in all likelihood reacted more quickly to new information than the average
investor, and while not necessarily creating an efficient market, it rather creates a market that is
much more prone to bullish and bearish trends over shorter periods of time; investors in the
cryptocurrency market have also tended to be more price elastic, which has led to more
overreaction than in a normal market which contributes towards the relative volatility in the
market. A 7-day rolling window is most likely to take this information into account most
effectively as it can quickly identify when a bull market is occurring and update efficiently,
while also significantly lowering the recommended allocation once the market begins turning
more bearish.
The general Modified Kelly strategy offers returns which at present are the same as those earned
through a passive strategy – as it essentially morphs into a passive strategy through always
recommending a full allocation into the asset when leverage is ignored. As the recommended
Kelly allocation continues to decrease for all cryptocurrencies, trending towards a 1 and
potentially lower allocation, more data may contribute to a more robust trading strategy that can
be utilised with this method, but more data will need to be accumulated first. This suggests that
more data is necessary over longer durations. As MacLean, Thorp, and Ziemba (2011) note,
Kelly is a long-term growth optimisation strategy and is most useful when considered over an
infinite period. Therefore, readers should be aware that the results obtained in this study can be
said to be focused heavily around the single market boom and bust, which may be considered an
anomaly in the long-term purview of the cryptocurrency market. With more extensive data,
especially if the market continues to normalise relative to its previous constant bullish states, the
normal modified Kelly recommendations for allocation should begin to recommend an allocation
of less than 1 and as a result may alter wealth allocations into fractions and not recommend
investing fully in the asset over entire periods.

There are three main factors which have been neglected but would alter returns had they been
included in this study – taxes, transaction costs, and leveraging. As cryptocurrencies have
generally been classified as an asset by governments, the returns gained can incur capital gains
taxes. In this study, because the asset in consideration is cryptocurrency, the transactions costs
associated with them are relatively lower. However, this may not be the case in the long-term for
some strategies because, for instance the modified Kelly using rolling periods with expected
weighted moving averages suggests a daily rebalancing of the portfolio, meaning trades would
be initiated daily. These transaction costs, although notably minimal on a relative basis (between

25
30 and 100 basis points per trade according to Gui, Hortacsu & Tudon (2018)) would tend to
accumulate and reduce the amount of return generated. It is believed that leveraging would be
likely to amplify returns significantly. In this study, there were many instances where a positive
return was made and the Kelly fraction suggested an allocation greater than 1, but the bounds of
the study capped the Kelly allocation to 1. This implies investors could have gained returns
through leveraging, but with the changes in the market that have been occurring, it would be a
mistake to recommend significant frequent leveraging until more trends in the post-boom
cryptocurrency market come to light.

Finally, each investor has his/her own risk tolerance level with which they are willing to engage
in before avoiding an investment. This study demonstrates that instead of following the full
modified Kelly using a 7-day rolling period with exponential weighted moving average, the
investor can optimise the growth of their wealth, through optimising their utility. Exponential
utility as tested in this study is not recommended for risk-averse investor usage for mitigation of
risk due to the lack of significant convergence, as well as lack of significant return relative to
movement in the cryptocurrencies. Risk-averse investors in general are likely to and likely
should avoid cryptocurrencies as a whole due largely to the overall volatile nature of the market.

For risk-seeking investors, while half-Kelly and quarter-Kelly strategies along with other similar
linear risk mitigating strategies that are used by relatively more risk-averse investors are
certainly not fully adequate to approximate what proportion of a recommended Kelly allocation
should be invested in order to mitigate enough risk while still somewhat maximising return, they
may be the best strategies currently available for those investors who still wish to utilise a Kelly
strategy to some extent. Tests on exponential utility for the post-boom period resulted in
similarly lackluster results, and as such the best possible recommendation for it is inconclusive.
As the cryptocurrency market expands and develops, there are numerous opportunities for
investors to apply different strategies and profit off of trends in the market. This paper focused
mainly on the modified Kelly strategy, using rolling windows along with exponentially weighted
moving averages while including short selling if it increased potential to magnify returns. An
exponential utility was tested, but ultimately failed as an airtight way to mitigate for potential
Kelly investors who may be more risk-averse, and even mostly for those who are more risk-
seeking especially relative to risk-adjusted returns available from 7-day EWMA windows. Thus,
future research could expand on comparing Kelly to different strategies in the cryptocurrency
market, perhaps attempting to use another relevant utility function to mitigate investor risk or
trying to incorporate leveraging. Furthermore, a more practical concise analysis of transaction
costs and how it would affect returns is an area that could be explored as many studies, including
this one, ignore transaction costs in their assumptions. Ultimately, the best way to incorporate a
Kelly strategy into the market might be to wait for more data to become available in the newer,
more normalised market and continuing to re-test along the lines used in this paper.

26
6. Conclusion
This paper aimed to establish whether a Kelly criterion relevant to the cryptocurrency market
could be developed, and if so, in what way it could be most applied to the market to maximise
the growth of investor wealth over time.

A traditional Kelly criterion was tested against this paper’s novel developed modified Kelly
criterion, as both were applied to Bitcoin, Ethereum and Litecoin and compared against a
benchmark, a passive srategy of buying and holding the underlying cryptocurrencies. By
applying exponentially weighting moving averages for a 7-day, 30-day, 60-day and 90-day
period, the resulting Kelly portfolios consistently significantly outperformed the passive strategy.
It could then be concluded that there is value to be added from an actively managed Kelly
strategy. The tendency for shorter weighting periods to outperform longer weighting periods
showed that for an investor to maximise their wealth over time, then quick reaction to market
sentiment is necessary. The Kelly fraction was also capped at 1 so the effects of gearing could
not be tested. For risk seeking investors further studies into this might be of value. Allowing for
short selling produced mixed results. It can then be concluded that access to short selling markets
is not a necessary part of maximizing wealth growth, though with the amplified results when
included with EWMA it can be said to be part of a wealth maximisation strategy. The omission
of transaction costs means that returns were overstated. However, looking at the 7-day
exponentially weighted moving average portfolio, the level of outperformance warrants the extra
costs of daily rebalancing. Based on the recommended Kelly allocations and subsequent results,
leveraging may also be viable during significant bullish and bearish periods.

7. References
Bloomberg, L.P. 2018. Price of SPDR S&P 500 ETF Trust Fund from 1 January 2013 to 31
August 2018 [Raw data]. Available: Bloomberg database [2018, October 5].

Board of Governors of the Federal Reserve System (US). 2018. 1-Month Treasury Constant
Maturity Rate [DGS1MO]. Available: https://fred.stlouisfed.org/series/DGS1MO [2018,
September 15].

Browne, S. 1995. Optimal investment policies for a firm with a random risk process: exponential
utility and minimizing the problem of ruin. Mathematics of Operations Research.
20(4):937-958.

Chen, S. & Huang, Y. 2008. Risk preference, forecasting accuracy and survival dynamics:
Simulations based on a multi-asset agent-based artificial stock market. Journal of
Economic Behavior & Organization, 67(3-4):702-717.

Chu, D., Wu, Y. & Swartz, T. 2018. Modified Kelly criteria. Journal of Quantitative Analysis in
Sports. 14(1):1-11.

27
Cocco, L. & Marchesi, M. 2016. Modeling and simulation of the economics of mining in the
Bitcoin market. PLOS ONE. 11(10): e0164603.
Coinmarketcap.com. 2018. Bitcoin Historical Data [Raw data]. Available:
https://coinmarketcap.com/currencies/bitcoin/historical-data/ [2018, September 1].
Coinmarketcap.com. 2018. Ethereum Historical Data [Raw data]. Available:
https://coinmarketcap.com/currencies/ethereum/historical-data/ [2018, September 1].
Coinmarketcap.com. 2018. Litecoin Historical Data [Raw data]. Available:
https://coinmarketcap.com/currencies/bitcoin/historical-data/ [2018, September 1].
Davis, M. & Lleo, S. 2011. Fractional Kelly strategies for benchmarked asset management. In The
Kelly Capital Growth Investment Criterion: Theory and Practice. 385-407.

Delbaen, F., Grandits, P., Rheinlander, T., Samperi, D., Schweizer, M. & Stricker, C. 2002.
Exponential hedging and entropic penalties. Mathematical Finance. 12(2):99-123.

Dyhrberg, A., Foley, S. & Svec, J. (2018). How investible is Bitcoin? Analyzing the liquidity and
transaction costs of Bitcoin markets. Economics Letters. 171(1):140-143.

Eeckhoudt, L., Gollier, C. & Schlesinger, H. 2005. Economic and financial decisions under
uncertainty. Princeton, USA: Princeton University Press.

Elendner, H., Trimborn, S., Ong, B. & Lee, T., 2016. The cross-section of crypto-currencies as
financial assets: an overview. SFB 649 Discussion Paper. 2016-2038.
Estrada, J. 2009. Geometric mean maximization: an overlooked portfolio approach?, The Journal
of Investing. 19(4):402-417.
Hagman, M.H. 2014. Money Management Using the Kelly Criterion. Master’s Thesis. UMEA
University.
Jiang, Z. & Liang, J. 2017. Cryptocurrency portfolio management with deep reinforcement
learning. In Intelligent Systems Conference (IntelliSys). 905-913.
Kelly, J. 1956. A new interpretation of information rate. IEEE Transactions on Information
Theory. 2(3):185-189.
Kim, G. & Shin, J. 2017. A comparison of the Kelly criterion and a mean-variance model to
portfolio selection with KOSPI 200. Industrial Engineering & Management Systems.
16(3):392-399.
Kuznitz, A., Kandel, S. & Fos, V. 2008. A portfolio choice model with utility from anticipation of
future consumption and stock market mean reversion. European Economic Review.
52(8):1338-1352.
Lo, A. W., Orr, A. H. & Zhang, R., 2018. The growth of relative wealth and the Kelly criterion.
Journal of Bioeconomics. 20(1):49-67.
MacLean, L., Thorp, E. & Ziemba, W., 2011. Good and bad properties of the Kelly criterion. In
The Kelly capital growth investment criterion: Theory and practice. 563-572.
28
MacLean, L., Thorp, E., Zhao, Y. & Ziemba, W. 2017. How does the Fortune’s Formula-Kelly
capital growth model perform?. Great Investment Ideas. World Scientific Publishing.
Mercereau, B. International Monetary Fund. 2003. Stock markets and the real exchange rate: an
intertemporal approach. Working paper. Available:
https://www.imf.org/en/Publications/WP/Issues/2016/12/30/Stock-Markets-and-the-Real-
Exchange-Rate-An-Intertemporal-Approach-16522
Narayanan, A. Miller, A., Felten, E., Bonneau, J. & Goldfeder, S. 2016. Bitcoin and
Cryptocurrency Technologies: A Comprehensive Introduction. Princeton, USA: Princeton
University Press.
Nekrasov, V. 2014. Kelly criterion for multivariate portfolios: A model-free approach. SSRN
Electronic Journal. 10(21):39-70.
Nielsen, M. 2014. The Kelly growth optimal strategy with a stop-loss rule. Quantitative Finance.
14(7):1131-1139.
Ohlsson, E. & Markusson, O. 2017. Application of the Kelly Criterion on a self-financing trading
portfolio-an empirical study on the Swedish stock market from 2005-2015. Bachelor’s
Thesis. University of Gothenburg.
Peterson, Z. 2018. The Kelly criterion in portfolio optimization: a decoupled problem. Journal of
Investment Strategies. 7(2): 53-76.
Raudys, A. & Pabarškaitė, Ž. 2018. Optimising the smoothness and accuracy of moving average
for stock price data. Technological and Economic Development of Economy. 24(3):984-
1003.
Roll, R. 1973. Evidence on the “growth-optimum” model. The Journal of Finance. 28(3):551-556.
Rouge, R. & El Karoui, N. 2000. Pricing via utility maximization and entropy. Mathematical
Finance. 10(2):259-276.
Rotando, L. & Thorp, E. 1992. The Kelly criterion and the stock market. The American
Mathematical Monthly. 99(10):922-931.
Thorp, E. 1962. Beat the Dealer: A Winning Strategy for the Game of Twenty One. New York:
Random House.
Thorp, E. 2011. Optimal gambling systems for favorable games. In The Kelly Capital Growth
Investment Criterion: Theory. Singapore: World Scientific Publishing.
Thorp, E. 1975. Portfolio choice and the Kelly criterion. Stochastic Optimization Models in
Finance. In The Kelly Capital Growth Investment Criterion. World Scientific Publishing.
Thorp, E. 2008. The Kelly criterion in blackjack sports betting, and the stock market. Handbook
of Asset and Liability Management. 5(1): 385-428.

29